Skip to main content

Progressive and adaptive learning for object manipulation: a biologically inspired multi-network architecture

Objective

We propose, in a collaborative effort between Neuroscientists and Roboticians, to develop and implement a biologically inspired multi-network control architecture for the Progressive and Adaptive Learning of Object MAnipulation (PALOMA) by anthropomorphic robots. This control system will incorporate architectural constraints for the interaction among multiple neural networks as defined by available neurobiological data on the connections between and within cortical areas. We will study the open-ended co-operation of successive learning stages for the progressive acquisition of multi-modal hand-object representations and multiple manipulatory tasks with increasing complexity. We will evaluate these learning capacities in two different, evolutionary robotic platforms integrating a hand, tactile sensors, and a vision system, to demonstrate through real-world interaction the feasibility, generality and adaptability of the learning scheme. We propose, in a collaborative effort between Neuroscientists and Roboticians, to develop and implement a biologically inspired multi-network control architecture for the Progressive and Adaptive Learning of Object MAnipulation (PALOMA) by anthropomorphic robots. This control system will incorporate architectural constraints for the interaction among multiple neural networks as defined by available neurobiological data on the connections between and within cortical areas. We will study the open-ended co-operation of successive learning stages for the progressive acquisition of multi-modal hand-object representations and multiple manipulatory tasks with increasing complexity. We will evaluate these learning capacities in two different, evolutionary robotic platforms integrating a hand, tactile sensors, and a vision system, to demonstrate through real-world interaction the feasibility, generality and adaptability of the learning scheme.

OBJECTIVES
The goal of our collaborative effort is to define, implement and demonstrate principles of a totally new neurobiologically inspired learning architecture for the sensori-motor control of robot manipulators.

First objective: to propose an innovative multi-network architecture, based on eurobiological knowledge of cortico-cortical neural networks, that enables progressive learning of multiple tasks.

Second objective: to model the underlying sensory signals used for adaptation and learning.

Third objective: to simulate five successive learning steps from simple sensori-motor interactions to manipulation.

Fourth objective: to demonstrate the feasibility and functionality of the novel learning scheme through embodiment in two different robotic platforms, which should provide evidence for adaptive object manipulation learned through real-world interaction.

DESCRIPTION OF WORK
The proposed group of 4 members consists of one laboratory dedicated to robot control (CR3, J Lopez-Coronado, CEDETEL), one laboratory dedicated to sensitisations (CR4, P Dario, SSSA) and two neuroscience laboratories. The latter are concerned with human based motor research on reach, grasp and manipulation (CR2, R Johansson, UMEA) and modelling of visuo-tactile-motor integration (CO1, Y Burnod, UPMC). We propose to develop a neurobiological inspired multi-network learning architecture for robot manipulators.

The development and implementation of the architecture will be divided into 3 stages:
1. Assessement: principles of robot vs. neural control. The goal is to assess existing robot control architectures (taken from the two robot labs) and the key elements and basic design principles of a neurobiologically inspired learning architecture (as defined by the two neuroscience labs), to define a multi-network architecture that enables progressive learning of multiple manipulation tasks of varying complexity;
2. Simulation: The goal of the second stage is the simulation of progressive, stage-wise learning by the use of the previously defined multi-network architecture. Five stages are envisioned for progressive learning: i) somato-motor stage, ii) visual and somato-motor stage, iii) reach and grasp of object, iv) sequence of movements, v) object manipulation;
3. Real-world demonstration: the overall objective of the proposal.

The goal is to implement the multi-network architecture in two real-world artefacts of different morphology and to demonstrate through real-world interaction the feasibility, functionality, adaptability and platform-independence of the proposed learning scheme. One platform (CR3, CEDETEL) is dedicated to industrial applications, the other (CR4, SSSA) to support for disabled and elderly people, and both include an anthropomorphic hand and a multi-modal perception system. Simulation and demonstration results will be u sed to revise the theoretical models.

Funding Scheme

CSC - Cost-sharing contracts

Coordinator

UNIVERSITE PARIS VI PIERRE ET MARIE CURIE
Address
Tour Centrale 21Eme Etage - 4 Pl Jussieu
75252 Paris Cedex 05
France

Participants (3)

CENTRO PARA EL DESARROLLO DE LAS TELECOMUNICACIONES DE CASTILLA Y LEON
Spain
Address
Parque Tecnologico De Boecillo P. 20B
47151 Boecillo (Valladolid)
SCUOLA SUPERIORE DI STUDI UNIVERSITARI E DI PERFEZIONAMENTO SANT'ANNA
Italy
Address
Piazza Martiri Della Liberta, 33
56127 Pisa
UMEA UNIVERSITET
Sweden
Address
C/o Ekonomi Adm Enheten
901 87 Umea